← Semantechs home Articles Feed

knowledge graphs · family history · RDF 1.2

Chapter and Verse

Knowledge Graphs for Family Historians and Investigators

· Peter Winstanley · 23 min read

On the first Tuesday of the month, the family history group meets in the village hall at the foot of the glen. The tables fill up with folders: printouts of census pages, copies of certificates, a photograph of a gravestone taken at an awkward angle, and, on one table, a family Bible with a cracked spine. Its owner has a problem that everyone in the room recognises. Her great-grandfather, Thomas Ashworth, was born in Halifax, but when? The flyleaf of the Bible says 1871. The 1911 census gives his age as 39, which makes it about 1872. And the baptism register of the parish church records his baptism on 3 March 1872, which says nothing about his birth except that it came first. Three records, three answers, and a family that has always said 1871.

Someone says what everyone in family history says sooner or later: "It's like being a detective." It is, and the questions on the tables show how much:

  • Which of the three dates is right, and why?
  • Is the Thomas Ashworth in the 1881 census the same Thomas as the one in the parish register, or another with the same name?
  • How exactly is she related to the Hartleys, and through which line?
  • Which links in the tree rest on a document, and which on a story told at a wedding?
  • If the Bible turns out to have been written up years after the events, what else in the tree depends on it?
  • And when the tree goes to a cousin in Canada who uses a different program, what arrives?

Change the parish register for a company filing, the family Bible for a witness statement and the cousin in Canada for a partner newsroom, and these are the questions of the first article in this series, written for investigators and analysts (Joining the Dots). That article argued that every published story has a bill of materials, whether or not anybody wrote it down. This one argues that family historians have been writing theirs down for longer than most, by hand, and have worked out more about evidence than they are usually given credit for. They have a written standard of proof. They have had a standard for exchanging their data since 1984. And the newest version of RDF, the standard behind knowledge graphs, now lets them record exactly what they have always meant: not "Thomas was born in 1871" but "the family Bible says Thomas was born in 1871".

This is the ninth article in a series on knowledge graphs. After investigators came the land question (Who Owns the Glen?), food (Field to Fork and Back Again), forests (The Wood and the Trees), electricity (Down the Line), a journey (Across the Water), a visit (Off the Beaten Track) and a town's year (Ebb and Flow). Each followed something through the hands of its many keepers. This one follows a family, and then returns to where the series began, with investigators, to see what they might learn from the people in the village hall.

A short history of family history

Family history runs on records that were never made for it. Parish registers of baptisms, marriages and burials go back to the sixteenth century. The state began registering births, marriages and deaths in England and Wales in 1837 and in Scotland in 1855 (nationalarchives.gov.uk). In Ireland it began with non-Catholic marriages in 1845, and covered everyone from 1864 (myheritage.com). Each register was kept for its own purpose: the church's sacraments, the state's population, the parish's rates. None of them was kept so that a great-granddaughter could find her way back, a century and a half later.

Ireland shows what happens when a keeper loses the records. Most of the census returns for 1821, 1831, 1841 and 1851 were destroyed in 1922, when the Public Record Office at the Four Courts was burned in the Civil War. The returns for 1861 to 1891 had already gone, destroyed by government order or pulped for paper during the First World War. The only complete returns to survive are those for 1901 and 1911, which had not yet been handed over (nationalarchives.ie). Irish family historians have spent a century working round that gap, with what they call census substitutes: tax valuations, church records, wills, newspapers. Every one of those is someone else's record, made for someone else's purpose, and has to be read with that in mind.

Two developments, half a century apart, changed the scale of the work. In 1894 the Genealogical Society of Utah was founded by The Church of Jesus Christ of Latter-day Saints, for whom family history is a religious duty. In 1938 it began microfilming records, and over the following decades filmed parish registers, civil records and archives in dozens of countries, on some 2.4 million rolls of film now kept in a vault cut into a granite mountain near Salt Lake City (thechurchnews.com). The society is now FamilySearch. And in 1984, when family historians were starting to use home computers, the same church published GEDCOM, short for Genealogical Data Communication, so that one family history program could pass a tree to another. Its version 5.5.1, drafted in 1999, became the format nearly every program reads and writes. Version 7.0 followed in 2021 (github.com). Few other communities of volunteers have had a shared data standard for forty years. Investigative journalism, which pools data across borders more and more, does not.

Evidence, not just records

Family history has also thought hard about what counts as proof, and written the answer down. In the United States, the Board for Certification of Genealogists sets out a Genealogical Proof Standard with five elements (en.wikipedia.org):

  • a reasonably exhaustive search for the records that might bear on the question;
  • complete and accurate citation of the sources used;
  • analysis and correlation of the information they contain;
  • resolution of any evidence that conflicts;
  • a soundly reasoned, written conclusion.

Read that list with an investigations editor's eyes. It is the bill of materials, with the inspection built in. A genealogist who says "Thomas was born about 1872" is expected to be able to show every record looked at, what each one says, why the Bible's 1871 was set aside, and the reasoning that ties it together. A conclusion that meets the standard is not certain. It is the best-supported answer on the evidence so far, open to revision when a new record turns up. That is exactly the stance an investigator takes, and it is the open world of the first article: the absence of a record is not evidence that the thing didn't happen.

The careful practitioners go further. They distinguish an original record from a copy of it, information from someone who was there from information passed on, and evidence that answers the question directly from evidence that only bears on it. A baptism entry made by the curate a week after the baptism is an original record, with first-hand information about the baptism and second-hand information about the birth. A date written into a Bible thirty years later, in the same ink as every other entry, is something else.

They also know which silences mean something. If a man is missing from the burial register of a parish that recorded every burial, and the register survives intact, his absence tells you something: perhaps he died elsewhere. Genealogists call that negative evidence. If the register is lost, or has gaps, the same absence tells you nothing at all. A family historian who writes "he did not die in Halifax" because a search found nothing has turned "not known" into "no", and a good reviewer will send the conclusion back.

FamilySearch built the same distinction into the data model it designed to succeed GEDCOM. GEDCOM X separates what each record says from what a researcher concludes: the people named in a record on one side, the researcher's conclusions about people on the other, and references from each conclusion to the evidence it rests on (gedcomx.org). The person in the parish register and the person in the 1881 census are two records. The researcher's Thomas Ashworth, born about 1872, is a conclusion that points to both. GEDCOM X never displaced GEDCOM in the programs people use at home, and FamilySearch itself went on to publish GEDCOM 7. The idea is the right one, and it is the idea the rest of this article is about.

Sentences about a family

As in the earlier articles, the basic parts of a knowledge graph are short sentences of subject, verb and object. A slice of the Ashworths might read:

Thomas Ashworth — was born —> about 1872

Thomas Ashworth — was born in —> Halifax, Yorkshire

Arthur Ashworth — has parent —> Thomas Ashworth

Thomas Ashworth — is married to —> Ellen Barraclough

The verbs come from schema.org (schema.org), the vocabulary the big search engines use to read web pages, which already has words for a person's given name, family name, birth date, parents and spouse. Organisations, events, documents and categories, which an investigation needs more than a family tree does, have their words in Semantic Arts' gist (github.com), used throughout this series.

The subtlety from the first article matters even more here. Each person gets an identifier of their own, not a name. Family history is full of John Smiths, of sons named after fathers, of a dead child's name given to the next one born. Deciding that the Thomas in the census is the Thomas in the register is a judgement, and in a knowledge graph it is recorded as a judgement, with its reasons, so that it can be revisited. The large shared trees online, which anyone may edit, are a running argument about exactly this: whether two records are one person, and who gets to say so.

Who said so? Statements about statements

The trouble with the sentences above is that they are too sure of themselves. "Thomas Ashworth was born about 1872" is a conclusion, but it sits in the graph looking exactly like a fact. What the family history group actually knows is something more like this:

The 1911 census — says that —> (Thomas Ashworth — was born —> about 1872)

The family Bible — says that —> (Thomas Ashworth — was born —> 1871)

The baptism register — says that —> (Thomas Ashworth — was born —> before 3 March 1872)

The group's conclusion — (Thomas Ashworth — was born —> about 1872) — rests on —> all three, with the reasoning written down

These are sentences about sentences. The first article mentioned that the next version of RDF, then close to completion at the W3C, would allow them. RDF 1.2 has been a Candidate Recommendation since April 2026, the stage at which the W3C invites implementations (w3.org). It does this with two new pieces. A triple term is a sentence used as a thing that other sentences can talk about. A reifier is something that stands for one or more of those sentences and carries what is said about them: here, where each came from.

The detail that matters most is easy to miss. In RDF 1.2, a sentence can be talked about without being asserted. The graph can hold "the Bible says Thomas was born in 1871" without holding "Thomas was born in 1871". The census's claim, the Bible's claim and the baptism register's claim all sit in the graph side by side, each with its source. Only the conclusion is asserted, and its reifier points to the three claims it weighed. That is GEDCOM X's division between records and conclusions, in a standard any RDF tool can read. It is also, in miniature, the Genealogical Proof Standard written as data. The examples for this article do exactly this, and a few lines of SPARQL then answer the questions from the village hall: which sources disagree with the conclusion, and what in the tree rests on the Bible. They are published with the series' technical companion (Show Your Working) as worked examples W23 to W30, each of which opens in the Turtle Editor Viewer with one link.

To show how this works in practice, there is Kinfold, a small family tree app on semantechs.co.uk (semantechs.co.uk). It is there as an example of how RDF 1.2 and a knowledge graph can be used in genealogy, not as a rival to the programs family historians already use. It runs in the browser and keeps the tree on your own device, stored as RDF and queried by HOLOS, an RDF 1.2 triplestore that runs inside the page. A birth, a death and each link to a parent or partner can carry their own sources, as RDF 1.2 reifiers. Correct a date in the app and its sources stay with it, in plain view, to be kept or changed. If a date is changed some other way, the sources recorded for the old one are kept and flagged, not lost. Two records of the same person can be linked as one, and the link is itself a statement with its evidence and its date, so it can be reviewed and, if it was wrong, undone; "not the same person" is recorded the same way, and saying nothing leaves the question open. Searches are SPARQL, built from a form, and a saved search is kept in the tree as linked data like everything else. The app does not yet keep the competing claims alongside the conclusion, as the examples do. That is the obvious next step, and RDF 1.2 makes it straightforward.

The question families ask most often is the one that gives Kinfold's relationship screen its name: how are we related? Answering it is a matter of finding a path through the graph and then naming it. Up two generations and down two is a first cousin. Up five and down six is a fourth cousin once removed. Up twenty-four is a twenty-second great-grandfather, and the same rules name him as easily as a grandmother. Marriage adds its own words: a stepmother is a parent's partner who is not a parent, and a co-father-in-law is your son's wife's father. Where no word exists, the plain chain is the answer: "your wife's brother's wife".

Two features of real families are worth a closer look, because investigators meet them too. The first is that people can be related in more than one way. When two brothers marry two sisters, their children are double first cousins, connected through both sets of grandparents. When cousins marry, which in many places and periods was common, their children descend from the same ancestors twice over, and family historians call it pedigree collapse. A good answer says so: the closest relationship, and the other line as well. RelFinder (semantechs.co.uk), another tool on semantechs.co.uk, draws every path between two or more things in a graph. Given a family tree, it shows the double cousins' two lines at a glance. Given a corporate register, it shows the director who is connected to the minister through three companies and a trust, by several routes at once.

The second is that the kind of link matters as much as the link. A child may have birth parents and adoptive parents, or foster parents. In the records of The Church of Jesus Christ of Latter-day Saints a child may be sealed to parents in a temple ceremony. GEDCOM has a field for this, which it calls the pedigree linkage type. A path through an adoption means something different from a path through a birth, just as a path through a nominee shareholder means something different from a path through the beneficial owner. With RDF 1.2 the kind of link, its dates and its sources can all be recorded on the link itself. And a path can then be judged by its weakest link: a relationship that runs through one well-sourced birth and one family story is only as good as the story.

Checking the tree

Every family tree program worth using has some version of a health check, and Kinfold's is typical. It looks for people who died before they were born, parents younger than twelve, parents who died more than a year before a child's birth, anyone with more than two parents, and the most embarrassing of all, someone who has become their own ancestor through a slip of the mouse. It also now finds sources recorded for a date or place that has since changed. These are the goods-in inspection of the first article, and they can be written as SHACL shapes, the W3C's language for saying what well-formed data must look like. The examples include the Genealogical Proof Standard's citation rule as a shape: every claim and every conclusion must point to what it was derived from. They check it with an open-source SHACL engine that, like HOLOS, runs in the browser.

The shapes say what must be present. They cannot say what is true. A check that flags a mother aged eleven is pointing at a probable mistake, not proving one. Records are wrong, transcriptions are wrong, and sometimes the improbable happened. The check tells you where to look, which is all a checker can do and as much as anyone should want.

Passing it on

The cousin in Canada brings us to GEDCOM, and to an experiment. To see how much of a family tree survives the journey from one program to another, I collected 43 public GEDCOM files and passed each one through Kinfold and back:

  • the "torture test" files created by H. Eichmann and extended around 2001 by J. A. Nairn, author of the GEDitCOM program, to put programs through every corner of the standard, accented characters included;
  • a set of awkward cases from the genealogy company Findmypast (github.com): same-sex marriages, children with no recorded parents, a man married to himself;
  • FamilySearch's own test files for GEDCOM 7 (gedcom.io), including remarriages and the Church's ordinance records;
  • example trees, from the Brontës to the Habsburgs, the largest with 34,020 people (github.com).

The first run was challenging. Old files written in ANSEL, the character set GEDCOM used before Unicode, came in garbled. A suffix such as "Jr." moved into the given name. A child with birth and adoptive parents came back with only one of them. Two parents who had never been a couple came back married. Dates written in words, such as "1076 or 1088", came back altered. Each of those was a few lines to put right, and after the fixes 39 of the 43 files make the round trip with every person, every link and every name intact, and exporting the same tree twice gives the same file.

What is still lost is more interesting than what was fixed, because it is lost by design: Kinfold has nowhere yet to put it. Marriage dates and places go, and so do divorces. The kind of each parent link goes, so an adopted child arrives with all four parents and no way to tell them apart, and the health check points at them. The Church's ordinance records go, as do godparents and married names. And brothers and sisters recorded with no parents at all lose their link, because Kinfold, like most programs, knows siblings only through the parents they share.

The lesson travels well beyond genealogy. Every exchange of data loses whatever the receiving model cannot say. What gets lost is almost never the people or the bare links. It is the qualifiers: what kind of link, when it held, who said so. Those are precisely what an investigation depends on. A knowledge graph helps in two ways. It can hold structure it does not yet understand as data, instead of dropping it at the door. And RDF 1.2 gives the qualifiers somewhere to live, on the link they qualify.

Where the two worlds met

Family history and investigation are not merely alike. In 2018 they met head on. Investigators in California uploaded a DNA profile from crime scenes of the 1970s and 1980s to GEDmatch, a free site where people compare their DNA to find relatives. It matched distant cousins. From those cousins the investigators built family trees, backwards to shared ancestors and forwards to their descendants, until they had men of the right age who had lived in the right places. In April 2018 they arrested Joseph James DeAngelo as the man known as the Golden State Killer (the-scientist.com). GEDmatch said it had not been approached about the case, and the questions it raised about relatives who had never consented to their DNA being searched have been argued over ever since.

What made the identification possible was not the DNA alone. It was a family historian's craft: building trees from fragments, weighing each record, correlating evidence, ruling people in and out. The method now has a name, investigative genetic genealogy, and its practitioners face the same discipline as the group in the village hall, with far higher stakes. Every link in a tree that points at a suspect had better be able to show its sources.

Who knew what, and when

There is another kind of investigation where the shape of the questions is the same and the stakes are higher still. A public inquiry sets out to establish what happened, why, and whether it could have been prevented. It hears evidence for months. Witnesses refine on Tuesday what they said on Monday. Some give evidence anonymously, because they raised concerns at work or for their own protection. Documents turn out to be incomplete. And the question at the centre is very often about knowledge over time: who knew what, when, and whether they passed it on.

To show how that looks in a knowledge graph without touching any real case, the second example accompanying this article is an invented inquiry. After a week of heavy rain, the earth embankment of a small hydro reservoir gives way in the night and floods the village below it. Nobody is killed, but dozens of homes are ruined, and an inquiry is asked whether the breach could have been prevented. The reservoir, the village, the company, the regulator, the council, the witnesses, the figures and the dates are all made up.

Even the first fact needs care. The council says 64 properties were flooded; the inquiry's own count is 58 homes. Both are right. The council counted every building with water in it, the village shop and the outbuildings included, and the inquiry counted the places where people live. A figure that does not say what it counts, and on whose definition, is a figure waiting to be misread.

The evidence has a familiar shape. On the twelfth day of hearings, the hydro company's operations manager tells the inquiry that readings of water seeping through the embankment had been rising since early October, and that the company told the regulator. On the thirteenth, he clarifies that the report was made by telephone, not in writing. On the seventeenth, the regulator's inspector says that no report was received. An employee who raised concerns inside the company gives evidence anonymously, as Employee K. And nobody can say whether the council's emergency planners were told: the company's records don't show it, and the council's mailbox for that October was partly lost when its email moved to a new system.

Every element of that is something the rest of this article has been about. Testimony is attributed to a person, given on a date, and sometimes refined the next day. A claim is not a finding: "a witness told the inquiry that the company told the regulator" is a different sentence from "the company told the regulator", and the difference is the inquiry's whole purpose. Some people may be referred to only in ways that cannot identify them. And the central question is about knowledge over time.

That last question is where the open world matters most. Ask whether one body passed information to another, and there are three honest answers, not two. A witness or a document may say that it was passed on. A witness or a document may say, definitively, that it was not. Or nobody can say either way: the question has been asked, the records have been searched, and they are silent. The third answer is not the second. A mailbox with no message in it may mean that none was sent, or that the warning went by telephone, or that the mailbox is incomplete, or that nobody kept one. Like the parish register with gaps, its silence counts only if the record would certainly have shown the thing. A system that quietly lets a missing record stand for "no" is manufacturing findings, and in a public inquiry there could hardly be a worse mistake. There is a fourth state too: a question nobody has yet asked. A knowledge graph can tell all four apart, provided it is built to.

In the example, the evidence reads like this:

Witness W, on day 12 — affirms that —> (the company — was aware of —> rising seepage), from 3 October

Employee K, on day 15 — affirms that —> (Employee K — raised concerns with —> the company), on 5 October

Witness W, on day 13 — affirms that —> (the company — informed —> the regulator), on 9 October, by telephone

Witness V, on day 17 — denies that —> (the company — informed —> the regulator)

A search of the council's emergency planning mailbox for October — finds nothing either way about —> (the company — informed —> the council), and notes that the mailbox is incomplete

Witness W's statement on day 13 — revises —> his statement on day 12

Employee K — is a protected witness, known by a code name and nothing else

Each claim is a reifier, unasserted, with two separate labels. Its stance says what it says about the proposition: affirms, denies or unknown. Its standing says how far it is accepted: alleged, disputed, corroborated or withdrawn. Whether the regulator passed anything on to the council appears nowhere, because nobody has asked; the graph doesn't pretend to know. Four queries then do the work:

  • Contradictions. Every proposition that one source affirms and another denies, with both kept side by side.
  • Who knew what, when. Every claim about awareness and information passing, in the order of the events it describes and with its sources attached. The unknowns are listed as unknown, not left out.
  • As of a date. What had been said by a given hearing day, the inquiry's version of the editor's "what did we know when we published?".
  • Open questions. Every proposition someone has asked about and nobody has answered, with what was searched.

The recall question works as before: if a statement is withdrawn, everything derived from it, including every revision of it, comes back in one list. Shapes guard the rest:

  • every claim has exactly one stance;
  • a yes or a no must come from an attributed source and have a standing, while an unknown must point to a recorded search, so a silent mailbox cannot be turned into a denial;
  • a protected person carries a code name and nothing that could identify them.

HOLOS can go further, with an access policy that decides at query time who may see what.

A tool like this does not decide anything. It cannot tell you which witness is right, any more than it can tell the village hall which of Thomas's dates is true. What it can do is keep every claim with its source, keep the conclusions separate from the claims, and keep the reasoning that connects them where anyone entitled to check it can find it.

What others have learned

  • FamilySearch: records and conclusions are different things. GEDCOM X kept what each record says apart from what the researcher concluded, with references from one to the other (gedcomx.org).
  • Wikidata: keep the competing claims, and rank them. Every statement in Wikidata can carry references and qualifiers, and a rank: preferred, normal or deprecated. Conflicting values are kept, the best-supported is marked as preferred, and queries use it unless asked otherwise (wikidata.org).
  • OpenSanctions: every value has a source and a history. Each value is recorded with the dataset it came from and when it was first and last seen (opensanctions.org).
  • The genealogists: write the standard down. The Genealogical Proof Standard is short, public, and applied by volunteers every day. Few other fields that deal in evidence have anything so compact.
  • Ebb and Flow: two official figures for one plant. The eighth article in this series found two capacities for Dingle's treatment plant in two official records, 12,000 and 8,600 population equivalents, and neither explained the other (Ebb and Flow). A figure that does not carry what it counts, and on what basis, cannot be checked.

The lessons for family historians and investigators alike:

  • keep claims and conclusions apart, and link each conclusion to the claims it weighed;
  • record every claim with its source, who made it and when;
  • tell "yes", "no" and "not known" apart, record what was searched for each "not known", and never let a missing record stand for "no";
  • keep the claims you reject, marked as rejected, with the reason;
  • give each link its kind and its dates, not only its two ends;
  • say what each figure counts;
  • protect identities by design, with code names and rules that forbid anything else;
  • judge a path by its weakest link.

A proposal

The pieces for this already exist, and none of them needs a large budget. What I would propose, for family history societies and archives on one side and newsrooms, think tanks and inquiries on the other, has five elements:

  1. A small shared core. schema.org for people, as Kinfold uses; gist for organisations, events, documents and categories; RDF 1.2 reifiers for claims and conclusions; the W3C's PROV vocabulary for who said what and when, and for the searches behind each "not known"; and categories for a claim's stance (affirms, denies or unknown), its standing and a link's kind. Small enough to learn in an afternoon, and the same for a family tree as for an investigation.
  2. Tools that run on a laptop and keep the material there. Kinfold for family trees; RelFinder for paths; the Turtle Editor Viewer (semantechs.co.uk) for looking at the data and checking it against shapes; HOLOS and the SHACL engine underneath, in the browser. Nothing is uploaded, which matters as much to a family with a secret as to a journalist with a source.
  3. Rules as shapes. The Genealogical Proof Standard's citation rule, conflicts flagged instead of overwritten, protected identities enforced, each written once and applied to everything.
  4. Exchange that keeps the qualifiers. A bridge from GEDCOM to RDF 1.2 that keeps the kind of each parent link, the marriage and its dates, and every source at the level it was recorded. FollowTheMoney, the data model behind OpenSanctions, could be bridged the same way for investigations.
  5. Three pilots.
    • A family history society's transcription project published as claims: each transcribed entry a claim with its source, the transcriber and the date, so that a corrected transcription replaces nothing and explains itself.
    • Competing claims, evidence along the path, source recall and kinds of parent link in Kinfold, so that the group in the village hall can do on screen what they already do on paper.
    • A casebook for public-interest investigations on the same core: claims with stance, standing and dates, open questions kept as open, code names for protected people, and the recall question built in.

Back to the table

By nine o'clock the group is packing up. Thomas's three dates are still three dates. Nothing on the tables has changed, except that each date now has its source written beside it, the Bible's 1871 has a note saying why it was set aside, and the conclusion, about 1872, says which records it rests on and why. If the cousin in Canada turns up a fourth date, it will go in as a claim beside the others, and the conclusion will be weighed again, not overwritten. "Chapter and verse" is an old phrase for giving the exact source, down to the line. It is what the Genealogical Proof Standard asks of a family historian, what an editor asks of a reporter, and what an inquiry asks of a witness. The family tree on the landing and the detective's wall in the first article were always the same thing: people on cards, connections in string, and someone standing back to work out what it means. Knowledge graphs, with RDF 1.2, can now keep the chapter and verse for every piece of string.